J Braz Comput Soc (2013) 19:511–521 DOI 10.1007/s13173-013-0115-9 ORIGINAL PAPER Syntenic global alignment and its application to the gene prediction problem Said S. Adi · Carlos E. Ferreira Received: 27 February 2013 / Accepted: 6 June 2013 / Published online: 6 July 2013 © The Brazilian Computer Society 2013 Abstract Given the increasing number of available geno- 1 Introduction mic sequences, one now faces the task of identifying their protein coding regions. The gene prediction problem can be The gene prediction problem can be defined as the task addressed in several ways, and one of the most promising of finding the genes encoded in a genomic sequence of methods makes use of information derived from the com- interest. In other words, given a DNA sequence, we would parison of homologous sequences. In this work, we develop like to correctly pinpoint the start and end positions of the a new comparative-based gene prediction program, called exons that constitute one or all of its genes. Like the search Exon_Finder2. This tool is based on a new type of align- for promoters, CpG islands and other functional genomic ment we propose, called syntenic global alignment, that can regions, the search for genes, that takes place at the annota- deal satisfactorily with sequences that share regions with dif- tion phase of any genomic project, has undeniable practical ferent rates of conservation. In addition to this new type of importance. alignment itself, we also describe a dynamic programming In prokaryotic organisms, the task of gene finding seems to algorithm that computes a best syntenic global alignment be easier than in eukaryotics. In the former, most of the DNA of two sequences, as well as its related score. The applica- sequence is coding for protein. Furthermore, each prokary- bility of our approach was validated by the promising initial otic gene is a continuous stretch of coding bases, making results achieved by Exon_Finder2. On a benchmark includ- the identification of these regions a feasible task. The genes ing 120 pairs of human and mouse genomic sequences, most of most eukaryotic organisms, on the other hand, are sep- of their encoded genes were successfully identified by our arated by long stretches of intergenic DNA and their cod- program. ing fragments, called exons, are interrupted by non-coding ones, called introns. In addition to the exons and introns, the Keywords Sequences alignment · Dynamic programming · eukaryotic genes include a number of other elements, such as Gene prediction 5-UTR, 3-UTR and splicing (donor and acceptor) sites. The structure of a typical multi-exon eukaryotic gene is shown in Fig. 1. Gene prediction methods can be roughly classified into two main categories, called ab initio, or intrinsic, methods S. S. Adi (B) and similarity-based, or extrinsic, methods (see [17,30]for School of Computing, Federal University of Mato an extensive review on this topic). The first ones [1,12, Grosso do Sul (UFMS), CP 549, Campo Grande, 39,42,51] rely on statistical information that alone, or in MS 79070-900, Brazil e-mail: [email protected] conjunction with some signals previously identified in the DNA sequence, allows the identification of its coding, non- C. E. Ferreira coding and intergenic regions. Some intrinsic methods make Institute of Mathematics and Statistics (IME), use of Hidden Markov Models (HMMs) [7,25–28,45,50] University of São Paulo (USP), Rua do Matão 1010, Cidade Universitãria, São Paulo, SP 05508-900, Brazil in order to combine both signal and statistical information e-mail: [email protected] concerning the target genes. The similarity-based methods 123 512 J Braz Comput Soc (2013) 19:511–521 Fig. 1 Simplified structure of a multi-exon gene [9,13–15,21,23,38,41,53] make use of similarity informa- using a set of multiple parameters with different levels of tion between the genomic sequence and a fully annotated stringency. transcript sequence, such as cDNA, EST or protein, in order We propose in this work a new type of alignment, called to accomplish the gene prediction task. syntenic global alignment, jointly with an algorithm that, Recently, with the huge amount of newly sequenced given two sequences, constructs a best syntenic global align- genomes, new similarity-based methods are being success- ment between them and calculates the associated value of fully applied in the task of gene prediction. In some ways similarity. This alignment can be seen as a generalization different from traditional extrinsic methods, the so-called of the generalized global alignment where three types of comparative-based methods [5,10,32,35–37,49], pioneered blocks are taken into account, and the corresponding algo- by Batzoglou et al. [4] with Rosetta, rely on similari- rithm is a special case of that proposed by Huang and ties between regions of two or more unannotated genomic Brutlag. sequences in order to find the genes encoded in each of In order to evaluate the applicability of our approach, them. The main assumption of these methods is that the func- the proposed alignment algorithm was used in the develop- tional parts of the eukaryotic genomic sequences, the coding ment of a new gene prediction tool called Exon_Finder2. regions, tend to be more conserved than the non-functional Our program was tested on two different benchmarks that ones. Finally, it is important to make reference to gene pre- include several pairs of real human and mouse genomic diction tools that combine extrinsic and intrinsic informa- sequences. The first benchmark includes 50 pairs of genomic tion. This is the case, for example, of Augustus- PPX [22], sequences taken from two traditional datasets. The second Twinscan [24], DoubleScan [33] and GenomeScan [52]. benchmark includes 70 pairs of genomic sequences. These Despite the enormous progress made to date (see Brent and pairs were obtained by us taking as base the human chro- Guigó [6] and Sleator [43] for a survey on this topic), the mosome sequences of the ENCODE project and their cor- gene identification problem remains an interesting subject of responding annotation. The genes encoded in a number of research. sequences that constitute these benchmarks were correctly Given the importance of genome comparison in obtaining located by our approach. information about these types of data, a number of heuris- This paper is organized as follows. In the next section tics algorithms aimed at constructing biologically meaning- we introduce the syntenic global alignment and show the ful alignments were developed [3,18,29,31,47]. In order to recurrences that allow us to find an optimal alignment of this deal specifically with sequences whose conserved regions type. Details about the use of this algorithm as a tool to the are intervened by unconserved ones, such as protein and gene prediction task are given in Sect. 2.1. The experimental prokaryotic gene sequences, Huang and Chao proposed in results are shown in Sect. 3. In the final section we make [20] the generalized global alignment. This type of alignment some concluding remarks concerning this work. discriminates between conserved and unconserved regions by using the concept of difference blocks. Unfortunately, there are still situations where even the generalized global 2 Syntenic global alignment alignment cannot be applied in a meaningful way. This hap- pens, for example, when the sequences to be compared Despite their practical importance, traditional alignment include highly conserved regions intervened by conserved algorithms cannot be used directly in aligning two genomic and unconserved ones. This is exactly the case in stretches sequences that share a number of strongly similar regions of eukaryotic genomic sequences that encode one or more intervened by regions with a low degree of similarity. genes. With the practical restrictions of the generalized global When sequences with these features are taken as input, the alignment, Huang and Brutlag describe in [19] an algorithm Needleman–Wunsch [34] alignment algorithm tends to align that computes an optimal alignment of two sequences by even unrelated regions (global alignment). On the other hand, 123 J Braz Comput Soc (2013) 19:511–521 513 the Smith–Waterman [44] algorithm identifies only a high- scoring similar region shared by the sequences (local align- ment). In order to deal with sequences that have intermittent Fig. 2 Example of a syntenic global alignment similarities, Huang and Chao proposed in [20] a variant of the global alignment called generalized global alignment.In and subsequent spaces in a gap of length l > 1 inside a highly such work, the notion of difference block is introduced. Such conserved block (resp. conserved block). Finally, let d be a a block includes residues that fall inside unconserved regions real value corresponding to a cost of each unconserved block. of the sequences that are being compared. With this new Given these definitions, the score of a syntenic global align- block, the task is to search for a best alignment of the input ment A is the sum of the values of each match, mismatch, gap sequences allowing the use of gaps, matches, mismatches and and unconserved block in this alignment. With these defini- differences. To this end, the authors suggest a dynamic pro- tions in mind, the problem we consider is the following: given gramming algorithm that makes use of four different matri- two sequences X = x , x ,...,x and Y = y , y ,...,y , , , 1 2 m 1 2 n ces: S I D and H. The first one is related to matches and find an optimal syntenic global alignment of X and Y , that mismatches. The matrices I and D deal with indels when is, a syntenic global alignment of these sequences with a comparing the sequences.